What for?

A shitty experiment of Rectified Flow for SDXL you can find it here in [Civitai] (https://civitai.com/models/2071356/experimental-noobai-with-rectified-flow-eq-vae)

Hard requirements

• Torch_linear_assigment or SciPy
• A working venv, you can take the one from LoRa_Easy_Training_Scripts/backend/sd-scripts/venv and install whatever is missing.

Additional Commands

• --flow_model Required to train into Rectified Flow target
• --flow_use_ot Not needed, want info? ask lodestone
• --flow_timestep_distribution = uniform or logit_normal, just run uniform
• --flow_uniform_static_ratio allows static values for shift, default 2.5
• --contrastive_flow_matching Magic thingie that makes training results a bit sharper
• --cfm_lambda needed by the one above, default 0.05, I prefer 0.02
• --flow_logit_mean needed by logit_normal timestep_distribution
• --flow_logit_std needed by logit_normal timestep_distribution
• --flow_uniform_base_pixels default 1048576 or 1024x1024 allows dynamic shifting based on resolutions useful for higher than 1024x1024 training.
• --flow_uniform_shift allows dynamic shifting
• --vae_custom_scale suggested for Anzhc's eq-vae put 0.1406
• --vae_custom_shift suggested for Anzhc's eq-vae put -0.4743
• --vae_reflection_padding suggested to use with Anzhc's eq-vae, my shitty experiment wasn't trained with this.
• --use_sga attaches Lodestone's stochastic accumulator for gradient accumulation, currently borked.
• --vae_type flux2 Self Explanatory, can work with Flux 1 too
• --latent_channels 32 Self Explanatory, use 16 for Flux 1 if you into that.
• --skip_existing Skips latents check, imagine checking ++10 Million latents when it can take a couple hours and time is money.
• --vae_custom_scale 0.6043 Flux 2 VAE
• --vae_custom_shift 0.0760 Flux 2 VAE

You can use launch_all_train.sh to start training, or copy the command to an activated venv, works on windows or linux w/e the file itself has a working template

Info on the shitty test

Trained with --flow_use_ot --flow_timestep_distribution uniform --flow_uniform_static_ratio 2.5 AdamW8bit Kahan Summation at 2e-5, caption and tag dropout of 0.1 with keep token separator based on NoobAI's own dataset, for one combined epoch of 486k unique images, scheduler was constant with warmup with 10% warmup.

Dataset

Uses Deepghs' danbooru webp from 0000 to 0065 with the exact same tags as NoobAI's own run extra files not present in NoobAI's dataset were purged.

Why tho?

Just sharing the codebase used to train the Rectified Flow model, so both people that know can scrutinize it, and people that think they do can shit on it.

How to train?

Activate venv, install missing shit, edit config and dataset toml as per your needs run launch_all_training.sh, or run accelerate manually, simple as that.